verifiers 0.2.2.dev82__py3-none-any.whl → 0.2.2.dev84__py3-none-any.whl

verifiers/v1/graph.py

@@ -101,6 +101,11 @@ class MessageNode(BaseModel):
     logprobs: list[float] = Field(default_factory=list)
     """Sampling logprobs for the sampled tokens — length equals the number of True entries in
     `mask`; empty for input messages."""
+    advantages: list[float] | None = None
+    """Per-token credit over the sampled tokens, same layout as `logprobs`. `None` until a
+    consumer's RL algorithm assigns it, which is not the same as a credit of zero: a group whose
+    rewards were all equal is assigned zeros and carries no gradient, while an unassigned node was
+    never scored at all."""
     multi_modal_data: SkipJsonSchema[MultiModalData | None] = None
     """The renderer items for the images this message's content introduces (pixel tensors,
     grids, hashes, placeholders) — the only carrier of the pixels from the env server to the

verifiers/v1/trace.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 import time
 import traceback
 import uuid
-from collections.abc import Mapping
+from collections.abc import Callable, Mapping
 from typing import TYPE_CHECKING, Annotated, Any, Generic, Literal
 
 import numpy as np
@@ -200,28 +200,45 @@ class Branch(BaseModel):
             mask.extend(node.mask)
         return mask
 
-    @property
-    def logprobs(self) -> list[float]:
-        """Per-token sampling logprobs aligned to `token_ids` — the node logprobs spread onto
-        their sampled positions, 0.0 on every non-sampled token."""
+    def spread(
+        self, values: Callable[[MessageNode], list[float] | None]
+    ) -> list[float]:
+        """A per-sampled-token node field widened to `token_ids`: each node's values land on its
+        sampled positions, 0.0 everywhere else. A node holding nothing contributes zeros."""
         out: list[float] = []
         for node in self.nodes:
-            mask = node.mask
-            sampled = sum(mask) if node.logprobs else 0
+            mask, node_values = node.mask, values(node) or []
+            sampled = sum(mask) if node_values else 0
             # Bulk-fill the canonical unsampled-prefix/sampled-suffix layout.
             if not sampled or all(mask[-sampled:]):
-                out += [0.0] * (len(mask) - sampled) + node.logprobs[:sampled]
-                out += [0.0] * max(0, sampled - len(node.logprobs))
+                out += [0.0] * (len(mask) - sampled) + node_values[:sampled]
+                out += [0.0] * max(0, sampled - len(node_values))
                 continue
             li = 0
-            for sampled in mask:
-                if sampled:
-                    out.append(node.logprobs[li] if li < len(node.logprobs) else 0.0)
+            for is_sampled in mask:
+                if is_sampled:
+                    out.append(node_values[li] if li < len(node_values) else 0.0)
                     li += 1
                 else:
                     out.append(0.0)
         return out
 
+    @property
+    def logprobs(self) -> list[float]:
+        """Per-token sampling logprobs aligned to `token_ids` — the node logprobs spread onto
+        their sampled positions, 0.0 on every non-sampled token."""
+        return self.spread(lambda node: node.logprobs)
+
+    @property
+    def advantages(self) -> list[float] | None:
+        """Per-token credit aligned to `token_ids`, spread like `logprobs` — or `None` when no node
+        on the path was ever assigned any, which a branch of zeros would otherwise be
+        indistinguishable from. A partially assigned path spreads, the unassigned nodes reading
+        0.0."""
+        if all(node.advantages is None for node in self.nodes):
+            return None
+        return self.spread(lambda node: node.advantages)
+
     @property
     def multi_modal_data(self) -> MultiModalData | None:
         """Node image data concatenated in token order for training; never persisted."""

{verifiers-0.2.2.dev82.dist-info → verifiers-0.2.2.dev84.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev82
+Version: 0.2.2.dev84
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.2.2.dev82.dist-info → verifiers-0.2.2.dev84.dist-info}/RECORD

@@ -172,7 +172,7 @@ verifiers/v1/agent.py,sha256=MQVKGjUhrP8uzRXAq94uJHkEsdHFziPxQHZ7TOD324I,31306
 verifiers/v1/env.py,sha256=g6fpoG-Z9lM3vTKrrU-QRJY3S88_y9zsPHwBvfs5vdY,17865
 verifiers/v1/episode.py,sha256=8G5OaGOku9S_gavQdpfUZgRZfrf1zmpjFrBL-TsppIk,4123
 verifiers/v1/errors.py,sha256=kQeEPX06TwAIuiz7MlWFYirrpKbx5STi6A5b_dtQQoA,6885
-verifiers/v1/graph.py,sha256=aBAzh1Ibp3klKNYjvzNyJSbjV8Oi5TJ791dN7o6ex1E,29070
+verifiers/v1/graph.py,sha256=psE4gww42TuSy_B4rJfGauycGTzPes6Z2SKFAzPYg7M,29427
 verifiers/v1/harness.py,sha256=eKuqvYbQjnCMavkiYWZQDHJgH08m6SJ4PuevPuShjS0,11097
 verifiers/v1/judge.py,sha256=Pvr0C41ah1qkNSZ0WXdkDvBMfv5YSfY1REHncwlJ8cg,9346
 verifiers/v1/legacy.py,sha256=YvcWF6d8xvkeqUPM36oif0ZrjmB0mj7nNrfwha0Nr2w,22850
@@ -181,7 +181,7 @@ verifiers/v1/session.py,sha256=p8vz89DJUmb9r8dkq80zn2WBEqo8Vwh8ot1ZXm_WVAM,7210
 verifiers/v1/state.py,sha256=EckF2bWp-vV4b1jYJ9sLI5xrfGuI5spIgYYwW926toI,595
 verifiers/v1/task.py,sha256=jwMKiKlMtksTd8j2dcDlFtb6LC8BRE-N5jkNXlMp2jc,9457
 verifiers/v1/taskset.py,sha256=fp2E0IEhL_Ybj9cegZwljfTmW27p_30zTWHFKw_kXE4,4374
-verifiers/v1/trace.py,sha256=esJE1aT9jRCbBad-BVR9xjX5Emk4mHdYmC7Zct-3zK8,19129
+verifiers/v1/trace.py,sha256=Mjc9NaqmUVKXhn6RKaTaGVxPRldCGqgHxb74cqtp5_k,20026
 verifiers/v1/types.py,sha256=1PxamJspmoTc3OlFZAwH6o_2i_Tg9ZNvHxttIHWIxq8,8948
 verifiers/v1/acp/__init__.py,sha256=9SYCFtzGUM_wH98ldpVLcFhoq2M999jbG0tU5ODY70U,2297
 verifiers/v1/acp/_runner.py,sha256=BcYaNZHuawzCgxOVdhiF6PY_B1HMxIA1MwHy0mOzhSk,7740
@@ -337,8 +337,8 @@ verifiers/v1/utils/platform.py,sha256=56Ixmk1SER6q5LvzyYcA0hgzGpzhxKZL8Hiqp4c_XV
 verifiers/v1/utils/retries.py,sha256=Y2ZgrAjn-qNkRKZP_RVNL_7EK0iaRcZeCa404lpGYi4,5417
 verifiers/v1/utils/score.py,sha256=-R5Cog14r_tJ6Q_oOfGr5VPAm2CCUhofYZB6B9lm4wM,5787
 verifiers/v1/utils/version.py,sha256=-obEo_-l9-D8FLef4hYxncOe-uJpxrM1g2Hig_37Sgs,1607
-verifiers-0.2.2.dev82.dist-info/METADATA,sha256=hwKzS7edm6kC2DKpuNCPpith5Nc6wcowpC3UCtG6eqc,4545
-verifiers-0.2.2.dev82.dist-info/WHEEL,sha256=lCkmxWfQsSc9CfIClYeavTdQeEX2toPqufh9gI35EQA,87
-verifiers-0.2.2.dev82.dist-info/entry_points.txt,sha256=v6v0QT9vVExnfn4br42MostI_o3f3ZbWYlTormz-U3g,515
-verifiers-0.2.2.dev82.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.2.2.dev82.dist-info/RECORD,,
+verifiers-0.2.2.dev84.dist-info/METADATA,sha256=hu_D0aS0L0VjJxeOvodUkjLmh1Ag4g4VU7iMaD8Yo2g,4545
+verifiers-0.2.2.dev84.dist-info/WHEEL,sha256=lCkmxWfQsSc9CfIClYeavTdQeEX2toPqufh9gI35EQA,87
+verifiers-0.2.2.dev84.dist-info/entry_points.txt,sha256=v6v0QT9vVExnfn4br42MostI_o3f3ZbWYlTormz-U3g,515
+verifiers-0.2.2.dev84.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.2.2.dev84.dist-info/RECORD,,